In briefGRAIL: Explaining time-series models through meaningful intervals, not single timesteps
SummaryThis paper proposes GRAIL, a framework for explaining time-series model predictions. Instead of scoring individual timesteps, it uses domain-specific intervals, such as the P-wave and QRS complex in ECG beats, as players described by human-readable features. It generates on-manifold counterfactual baselines and attributes predictions with Owen values, and can also weigh static covariates like age against the signal. According to the paper, experiments on synthetic and real data show faithful and robust explanations useful to domain experts.
In briefGate Setup vs. Recurrent Regularisation for Long-Range Learning in LSTMs and GRUs
SummaryThis arXiv paper tests how gate initialisation (conventional versus chrono) and recurrent dropout or zoneout placement affect whether LSTM and GRU networks learn long dependencies, using 720 controlled runs on the adding problem and delayed recall. All chrono-initialised runs reached the validation criterion within 5000 updates, while conventionally initialised networks learned slowly or failed at longer lengths. The authors conclude the results support chrono initialisation on these synthetic tasks..
In briefPHD-Hand: A Low-Cost, 3D-Printable Humanoid Hand with Linkage-Driven Digits
SummaryThis paper presents a linkage-driven humanoid robot hand in which each digit has three degrees of actuation. It uses a unique flexure hinge and spring-steel elastic drive rods to simplify the drivetrain. The design avoids many miniature moving parts, so it can be fabricated with a budget 3D printer and sized closer to a human hand. The authors evaluate a prototype on grasping and object manipulation and report findings on its kinematics and control.
Neel Adwani, Muhaiminul Islam Akash, Saffia Ouali et al.
In briefBetter policies can make verifier audits worse
SummaryThe paper studies a control loop in which a component verifier is audited on fresh policy outputs against the full success label and refreshed when the score drops. As the policy masters the component, its own failure cases lose weight, so the audit can prefer a worse verifier, which the authors call on-policy audit collapse. They derive theoretical bounds and a repair: audit the component on its own failures and keep those failures in the refresh distribution.
WhyIdentifies on-policy audit collapse with sharp identified sets and erasure thresholds, validated by exact-oracle, text, and ECG experiments.
In briefResonance and Depth: Timbre and Space in an AI-Assisted Percussion Performance
SummaryThis artistic research report presents a 9-minute immersive electroacoustic percussion performance in which timbre and space serve as the main compositional materials. It is built from live electronic percussion and the author's own recorded corpus, processed by a multi-agent system that recombines and spatialises the material in real time. The report examines how performer identity can be retained, transformed, and redistributed across human, algorithmic, timbral, and spatial processes.
In briefSingle-photo food weight estimators fail to carry their gram scale across datasets
SummaryThe study fine-tunes DINOv2-base on 2,253 weighed Nutrition5k dishes to predict food mass from one RGB photo, and compares it with the zero-shot vision-language model Qwen3-VL-32B. On Nutrition5k and three external datasets, the fine-tuned model performs well in domain but does not beat Qwen externally, and its gram scale is inconsistent across scenes and foods. Adding one labeled photo of the same food reduces the ASA24 error.
In briefDissimilarity-Based MDPs for Stabilizing State Distributions
SummaryThis paper proposes a theoretical framework for learning policies that stabilize the state distribution of a stochastic system, not just maximize expected reward. It formulates the problem as a Dissimilarity-based Markov Decision Process (DMDP), whose objective measures the dissimilarity between the current and a target state distribution, and gives sufficient conditions under which the optimal policy makes the target distribution asymptotically stable. The authors also present a computable linear-system example and experiments on a HalfCheetah task, combining the approach with PPO and SAC.
Xun Shen, Yuepeng Wang, Daniyal Sajedi-Hosseini et al.
In briefInheritance as an admission test for LLM prompt merge operators
SummaryThe paper proposes an admission test for LLM prompt merge operators: a merge is accepted only if the child moves toward the donor beyond a zero-dose control, with the lower confidence bound reaching half the requested dose. On one negotiation task, tactic append was admitted, free rewrite was rejected, and template append was left undetermined. Results are measured with gemini-3.1-flash-lite.
In briefGenXR-Bench: Benchmarking Coding Agents for Interactive XR Development
SummaryThis paper introduces GenXR-Bench, the first benchmark for end-to-end evaluation of coding agents building interactive extended reality (XR) applications. It contains 112 expert-authored tasks covering feature implementation, bug repair, and application generation, tested with scripted and adaptive interactions. The strongest of 10 evaluated models reaches only 50.9% overall Pass@1. Failure analysis points to recurring misunderstandings in asynchronous sensor lifecycles, physical grounding, and embodied interaction semantics.
William Huang, David Li, Xiang 'Anthony' Chen et al.
In briefTarget regeneration shrinks gains from draft-training objectives in speculative decoding
SummaryThis paper studies training objectives for the draft model used in speculative decoding, a method that speeds up language models by verifying drafted tokens in parallel. The authors compare cross-entropy, token-level distillation, and distillation restricted to positions that survive verification. They find large gains on external responses, but after the target model greedily regenerates the data, distillation's gain is smaller and the gated objective gives little or no gain. The authors recommend measuring objective gains against a target-regenerated baseline and across training lengths.
WhyShows target regeneration shrinks draft-training gains, validated by controlled experiments across settings, corpora, and training lengths, offering practical guidance.
In briefEvidence-driven multi-agent AI for head and neck proton radiotherapy planning
SummaryThis paper presents a multi-agent compound AI system that combines deep reinforcement learning (a PPO-trained policy network) for automatic radiotherapy plan generation with an LLM-based conversational interface for refining plans. It was trained and evaluated on 21 head and neck cancer patients, with a post-trained Qwen3-14B planning agent using SFT and GRPO. Compared with the base LLM, the post-trained model showed better instruction-following, information retrieval and domain reasoning, improving user-prioritized objectives while limiting deterioration of competing ones.
In briefSparse, local kernel method for learning incompressible flow operators, using Lagrange weights
SummaryThe paper presents a local and sparse kernel method for learning solution operators of incompressible flows, extending the authors' earlier property-preserving kernel method. It replaces global output reconstruction with local divergence-free interpolants solved in parallel on GPUs, and uses a compactly supported kernel whose support radius is fitted to data to make the regression sparse. On ten fluid benchmarks, the sparse models are reported to be at least as accurate as the earlier dense method and more accurate than neural operators.
Ramansh Sharma, Matthew Lowery, Houman Owhadi et al.
In briefImproving Banana Variety Classification Across Quality Grades with Hybrid CNN-Transformers and Self-Distillation
SummaryThis paper addresses banana variety classification when models trained on some quality grades must generalize to unseen grades. Using a Leave-One-Grade-Out protocol on 31,678 images covering three varieties and three grades, it compares CNNs, ViTs, and dual-stream hybrid architectures. The proposed dual-stream Swin ResNet with KL-divergence-based self-distillation reaches a mean accuracy of 81.25%, outperforming CNN and Transformer baselines.
Khush Kataruka, Mohammad Muzammila, Zahid Akhtar et al.
In briefA jump-modified Wasserstein gradient flow for MMD with energy kernels
SummaryThe paper studies a jump-modified Wasserstein gradient flow of squared maximum mean discrepancy generated by energy kernels K(z)=-|z|^q with 0<q<2, in all dimensions d≥2 and in one dimension for 1≤q<2. Under subcritical density and finite-moment assumptions, the authors prove global well-posedness. Under their hypotheses on the target and jump rates, they show convergence to the target from finite-second-moment data.
WhyProposes a jump-modified MMD gradient flow with proofs of global well-posedness and convergence to the target; impact mainly confined to the theory subfield.
In briefVeriEvolve: Evaluator-First Evolution Framework for Tool-Using Agents
SummaryThe paper introduces VeriEvolve, which builds an executable evaluator from the task instruction before optimizing an agent, and uses it as a frozen fitness signal during workflow evolution. The evaluator is revised between rounds by turning end states of strong and weak runs into counterfactual test cases. On the 30 MCPMark Filesystem tasks, evaluator-guided evolution raised externally verified success from 6/30 to 18/30 for GLM-5.2 and from 27/30 to 30/30 for Claude.
In briefT-DIG: A Text-Conditioned Framework for Diverse Image Generation
SummaryThis paper presents T-DIG, an iterative framework that lets users request varied image outputs along axes they specify in natural language, such as background or artistic style. At its core, a language model policy writes multiple axis-conditioned prompts or edit instructions for a frozen image generator or editor, trained with GRPO. The authors also release the first dataset for text-conditioned prompt diversity, and report that T-DIG improves diversity over baselines while preserving alignment with the base prompt.
Simran Khanuja, Rohit Gandikota, Gaurav Parmar et al.
In briefFlashNext-RL: System optimization for long-horizon agentic reinforcement learning on Qwen3.8-Flash-Next
SummaryThe paper presents FlashNext-RL, a model-specific reinforcement learning training framework built on prime-rl, with vLLM used for software engineering rollout generation. It focuses on BF16 training of the full 48-layer model on a single 8xB200 node, optimizing host offload, gradient accumulation, and kernels for mixture-of-experts, sparse attention, and the gated residual stream. The authors report training throughput rising from about 395 to 4,438 tokens/s/GPU, and a live software engineering RL run averaging 22.4 minutes per iteration over steps 5-15.
In briefRethinking Time Series Tokenization Through Frequency Analysis
SummaryThe paper argues that common methods of splitting time series into subsequences for tokenization introduce spectral distortions that push learned representations away from true trends, periodicity, and seasonality, harming model performance. Using a frequency-domain analysis, the authors derive the boundary conditions under which these distortions arise. They propose TimeFT, a nearly distortion-free, parameter-free frequency-based tokenizer that can replace existing tokenizers. Experiments on forecasting, classification, and anomaly detection report consistent and significant performance gains.
In briefAn Optimal Sampler for Finding Diverse Traveling Salesman Routes
SummaryThis paper studies the Diverse Traveling Salesman Problem, which seeks k distinct tours that are each within a set multiple of the optimal length while keeping their pairwise similarity low. The authors propose a hybrid method that splits the task into a sampling stage and a selection stage, and they give theoretical guarantees for the optimal sampling distribution. Experiments report that a single trained model outperforms baselines on both in-distribution and out-of-distribution instances without fine-tuning.
In briefUsing aphasia as a lens to study how large language models store language functions
SummaryThis paper introduces a mechanistic interpretability framework inspired by aphasia, the language impairment caused by brain damage in humans. It uses two methods, classifier-guided lesioning and continuous activation steering, to induce aphasia-like breakdowns in LLMs. The authors find that language functions can be isolated to discrete circuits and continuous vectors, but these remain densely entangled, with substantial overlap across different aphasic profiles. Steering generalizes robustly across downstream tasks, while structural lesioning is asymmetrically constrained.
In briefDoob-Based Generative Policy Optimization: A New Way to Fine-Tune Diffusion and Flow Robot Policies with RL
SummaryDiffusion and flow models are expressive policies for robots and sequential decision-making, but their action likelihoods are hard to compute, which blocks standard methods like PPO and GRPO. The paper proposes DoobPO, which uses Doob's h-transform for likelihood-free reinforcement learning fine-tuning. It first learns a lightweight ratio model for the desired policy update, then distills it into the generative policy via conditional drift matching. On MuJoCo continuous-control benchmarks, DoobPO reportedly performs strongly against competitive diffusion and flow policy baselines.
In briefClassifier-based detection of indirect prompt injection using attention signals
SummaryIndirect prompt injection hides malicious instructions in external data that a language model processes, threatening deployed LLM agents. The paper evaluates Attention Tracker, a white-box detector based on attention distribution, on the AgentDojo benchmark and finds its claimed dataset-agnostic detection does not hold, with near-chance performance across most model families. The authors propose AttnDA, which replaces the scalar threshold with a Fisher Linear Discriminant Analysis classifier, improving detection across four benchmarks, though transfer to AgentDojo remains challenging.
In briefMedBOE: Estimating Complete Patient Body Models from Incomplete Whole-Body Medical Images
SummaryWhole-body medical scans can miss parts of the body, such as arms or hands, because of scanner size or patient positioning. This paper presents MedBOE, a training-free method that fits a parametric human body template (SMPL) to incomplete scans and estimates the missing regions. The method was tested on two public datasets, AutoPET and HIT, and achieved the best overall volumetric and geometric performance among the methods evaluated. Body-habitus metrics derived from its outputs were within 2% error.
In briefBehavior Beyond Text: Evaluating Refusal Transfer Failure in LLM Agents
SummaryThis paper examines whether language models that refuse harmful requests in text also refuse them when acting as tool-calling agents. The authors built a benchmark of 2,304 matched prompts across four domains and tested five models using 20 forbidden actions. They found that many requests refused in text still led to unsafe tool calls, and that a refusal-related direction in the models only partly explains this gap. The authors conclude that evaluating tool-calling agents requires directly examining their actions, not just their text responses.
Ekagra Gupta, Raj Saha, Aniekan-Abasi Inyang et al.
In briefThe Butterfly Effect in LLM Social Simulations: Robustness Audits
SummaryThis paper examines how small implementation changes in LLM social simulations can propagate through agent interactions and alter collective outcomes. Using a repeated Prisoner's Dilemma and a social-media echo chamber across four LLM families, the authors test representation- and design-level perturbations. Changing only a matched persona's representation shifts two-agent cooperation by 75.2 percentage points for GPT-5.2, showing robustness cannot be inferred from one configuration. The authors introduce TRAILS and the Social Simulation Audit Card to systematically audit these dependencies.
In briefFrom Common Information to Canonical Correlation Analysis: Linking Information Theory, Signal Processing, and Machine Learning
SummaryThis tutorial brings together the main results and tools that information theory, statistics, and signal processing have developed for recovering a shared component from several jointly observed signals. It builds the information-theoretic foundations from first principles, then covers canonical correlation analysis (CCA) and its multi-view extension, including its use as a blind receiver in multiuser communications. It also addresses nonlinear CCA (kernel, deep, and probabilistic forms), conditions for identifying the common factor, and links to self-supervised learning.
In briefDoes In-Context RL Act as a Bayesian Posterior Sampler? An Empirical Study
SummaryThis paper studies in-context reinforcement learning (ICRL), where models adapt to new tasks at inference without parameter updates. It asks whether Decision-Pretrained Transformers (DPT) and related variants actually perform Bayesian posterior sampling in practice, rather than only under idealized theory. The authors define two necessary conditions for such a sampler and build statistical tests to check them. In Gaussian, Bernoulli, and linear bandit settings, both conditions were consistently violated, suggesting ICRL may not behave as a posterior sampler in practice.
In briefSSMCGM-Stream: A streaming model that predicts personalized blood glucose for diabetes care.
SummaryThis paper presents SSMCGM-Stream, a streaming state-space model that updates a personalized hidden state with each new glucose reading to forecast blood glucose in constant time. It offers interpretable forecasts via hidden-state attribution and supports counterfactual reasoning. The companion SSMCGM-Agent turns forecasts and explanations into risk-aware, natural-language recommendations through a guarded language-model interface. On the T1DEXI cohort, the model matches strong neural baselines while greatly reducing computational cost, enabling on-device decision support.
In briefHow Orthogonalization Fits Neural Network Hessian Structure: A Gradient Self Outer-Product Analysis
SummaryThis paper examines why the Muon optimizer, which orthogonalizes momentum, suits the Hessian structure of neural networks. The authors analyze the gradient self outer-product at initialization across four models: symmetric matrix factorization, deep linear networks, two-layer ReLU networks, and single-layer linear attention. They find that diagonal entries grow faster than off-diagonal ones at explicit rates as hidden dimensions increase, with off-diagonal entries vanishing entirely for attention query, key, and value weights. Simulations and language model experiments at initialization confirm the predicted scaling.
In briefTEMPL: A Transparent, Model-Free Method for Pseudo-Labeling Material CT Scans
SummaryThe paper presents a multi-stage method that creates polygon pseudo-labels directly from grayscale CT scans, using thresholding, morphological dilation, and a custom region-growing algorithm, without pre-trained models, labeled data, or a GPU. On a dendrite dataset it reached 63.19% IoU, compared with 10.74% for zero-shot SAM. On a silica sand specimen, the median grain equivalent diameter was 163.8 micrometers, close to the known sieve range. The method takes about 15 seconds per image on a standard CPU.
Ashiqur Rahman, Venkata Devesh Reddy Seethi, Austin Yunker et al.
In briefFrom Weather to Enhanced Weathering: Weather Models for Scaled Carbon Removal Measurement
SummaryEnhanced Rock Weathering (ERW) is a carbon removal method that could remove gigatons of CO2 per year, but measuring the removal currently requires dense soil sampling and lab analysis. The authors tested five weather foundation models as frozen encoders with lightweight decoders on 5,192 sequences from 264 fields in Brazil and the US, using leave-one-field-out validation. Their best model, TerraNova-ERW, correctly ordered measurements within unseen fields, but zero-shot prediction of a new field's mean removal is not yet feasible. Few-shot prediction using a few samples per field is supported and can cut sampling on the densest fields by 28-98%.
WhyFirst holistic evaluation of weather foundation models for enhanced rock weathering, with large-scale field validation across 264 fields and 5,192 chronosequences.
Sasankh Munukutla, Felipe Oviedo, Danrong Zhang et al.
In briefPrivacy-protected quantum machine learning for clinical and wearable health data, with unit-aware noise
SummaryThis paper asks whether quantum machine learning models keep more predictive utility than classical models under differential privacy when applied to clinical and wearable health data. The authors compare capacity-matched quantum, multilayer perceptron, and logistic regression models across three clinical cohorts and one wearable cohort, using five privacy budgets and five seeds. The quantum model shows a trend of better performance on the clinical cohorts but no difference survives Holm correction, so the authors report a trend rather than a quantum advantage. They also run inference on an IBM 156-qubit quantum processor, reporting mean absolute probability errors relative to simulation.
In briefRoPT: Robust 6-DoF Object Pose Tracking in Video
SummaryThis paper presents RoPT, a framework for tracking an object's 6-DoF pose in video, aiming to prevent divergence when objects move rapidly. It couples visual mask tracking, a kinematic factor graph with multi-hypothesis verification, and active re-localization into a closed loop for self-recovery. The authors report higher tracking survivability under rapid motion on challenging benchmarks and real-world robotic demonstrations.
In briefTransformers as generic neural networks: exact compilation and multi-agent attention clustering
SummaryThe paper shows that every Transformer block, including attention, LayerNorm and a GELU MLP, can be exactly compiled into a generic deep network built from affine maps, routing, square, sigmoid, inverse square root and GELU. Without activation-activation products, dense attention needs at least N² scalar nonlinear evaluations per head, and the authors show that keeping that operation type avoids this cost. The work also analyzes top-k attention tail handling and uses centroids to select keys first, reducing computed scores from N² to Nk. On DeiT-Tiny, MambaIRv2-Light and Pythia-70M, the compiled models match their references and work with standard post-training quantization.
In briefScan or Index? Evidence Coverage, Answer Quality, and Cost in Long-Document RAG
SummaryThis paper compares 11 retrieval configurations for long-document question answering, including lexical and dense search, hybrid reranking, and full-document scanning. It evaluates them on 396 QASPER questions and 80 questions from eight textbooks. The highest evidence coverage does not always yield the best answers, full scanning has higher latency on books, and reusable dense retrieval has lower projected API cost as queries accumulate. The authors argue that evidence retention, answer quality, and reuse cost are complementary criteria.
Juan Terven, Diana-Margarita C\'ordova-Esparza, Julio-Alejandro Romero-Gonz\'alez et al.
In briefSLoSa: Small, Low-Power, Safe AI Models for Robot Task Planning
SummaryThis paper presents SLoSa, a framework for training small, low-power robot task planners built on foundation models that are aware of safety constraints. It builds a safety-aware dataset using classical planners, a teacher LLM, and formal verifiers, then applies supervised fine-tuning and reinforcement learning guided by a verification-based reward machine. The method is evaluated on four PDDL domains, two household simulators, edge hardware, and a desktop robotic arm, reporting success rates and power draw.
In briefPair-free video generation turning coarse renders into realistic footage
SummaryThis paper addresses converting coarse animation, simulation, or game-rendered videos into photorealistic videos while preserving layout, camera motion, and dynamics. The authors train DINO-R2C on easily built real-to-coarse image pairs to map realistic frames to coarse frames in DINO feature space, then apply it to ordinary real videos to create pseudo-coarse videos. The C2R# model learns from these pseudo-coarse videos, so no aligned video pairs are required.
Gonzalo G\'omez-Nogales, Daohan Lu, Peiye Zhuang et al.
In briefHyperDrone: Training drone vision in simulation at millions of frames per second
SummaryHyperDrone is a quadrotor simulator built for very fast batch rendering, reported to be over 20 times faster than existing batch quadrotor simulators. Its ray-tracing renderer, HyperLux, runs fully on the GPU across several graphics backends and produces camera-style observations at adjustable fidelity. The authors trained a vision-based yaw-control policy purely in simulation and transferred it to a small Crazyflie quadrotor with a camera that runs entirely onboard.
In briefOnline open-world wafer defect discovery using hyperdimensional computing
SummaryConventional offline classifiers struggle to detect new wafer defect patterns that appear during continuous semiconductor manufacturing. The paper proposes WaferHDC, a memory-centric framework based on hyperdimensional computing that detects anomalies and groups recurring failures into compact hypervectors without labels or predefined class counts. It outperforms the strongest online baseline on the WM-811K dataset in adjusted Rand index and normalized mutual information.
In briefIs 2:4 Sparsity Still Worth It? Comparing Semi-Structured Sparsity and Quantization for LLM Inference
SummaryThis paper compares two ways to compress large language models for inference: 2:4 semi-structured sparsity and lowering quantization precision. Measured at equal bits and equal tensor-core tiers, across all six GEMM operands of a transformer layer on three GPU generations, the authors find that adding 2:4 sparsity is worse than dropping to a lower precision when one is available. At the lowest supported precision, 2:4 still speeds up linear layers by 1.3 to 1.7 times, but accuracy must be recovered through retraining.
In briefSelf-improving AI agents that imagine outcomes before acting
SummaryThis paper proposes a test-time framework for LLM agents that combines multi-branch predictive analysis with experience-based adaptation, inspired by MPPI control. At each step, the agent proposes several candidate actions and uses a world model to imagine their outcomes without executing them. Predicted outcomes are scored for task progress and risk, and MPPI-style weights guide the final decision. Experiments on ALFWorld and ScienceWorld across three backbones show strong overall performance.
In briefGuilt by Association: Semantic Leakage as a Driver of Associative Gender Bias
SummaryThis paper studies semantic leakage, where a concept in a prompt that is irrelevant to an entity still shifts a language model's output about that entity. The authors find the effect persists even when the concept is not directly attributed to the entity, suggesting a broader associative tendency. They extend this to gender bias, showing that gendered concepts merely present in a scene push outputs toward stereotypes to a similar extent as direct engagement. The authors argue this form of bias is overlooked by current evaluations.
In briefMulti-objective protein design by steering diffusion models at inference time
SummaryThis paper presents MODS, a framework for optimizing several protein properties at once by steering a diffusion model during inference. Each proposal uses randomly sampled scalarization weights and a surrogate ensemble member to guide generation, and surrogates are updated after each evaluation batch without retraining the generator. Tested on six protein optimization tasks, MODS achieves competitive final hypervolume with both diffusion models and substantially outperforms the best baseline on several tasks.
In briefMINER: Modeling Person-Level Variance for Re-identification
SummaryThis paper addresses clothes-changing person re-identification (CC-ReID), where a person's appearance changes significantly with different outfits. The authors propose MINER, a model-agnostic framework that suppresses clothing bias by contrasting hard examples and adaptively balancing intra- and inter-person variance. Experiments report state-of-the-art results on four CC-ReID benchmarks and top rankings in all three tracks of the TVRID RGB-Depth ReID challenge.
In briefStyleStream 2.0: A Faster, Controllable Streaming Voice Style Converter
SummaryThis paper presents StyleStream 2.0, a streaming system that changes a speaker's timbre, accent, and emotion to match a target voice while keeping the spoken words. It cuts end-to-end latency to about 520 ms on a consumer GPU, roughly twice as fast as version 1.0. A text-conditioned model over a compact style embedding supports designing voices from text prompts and editing a single attribute of an utterance. The authors report the strongest target style fidelity, with competitive intelligibility and preservation of unedited attributes.
In briefChecking Conformal Prediction for Language Models When Inference Settings Change
SummaryConformal prediction for language models is usually tested with the same inference settings used at calibration, yet deployment often changes settings like quantization or temperature for memory or throughput reasons. The paper formalizes this as inference-time model shift, a change in the conformal score distribution caused by inference configuration rather than the input-label distribution, and shows theoretically that it breaks split-conformal exchangeability. Across 8 language models, 4 finite-label benchmarks, and 12 inference configurations, matched evaluation can substantially overstate deployment-time coverage, with effects varying by model and dataset. The authors argue inference configuration should be reported as a primary experimental variable.
In briefBack Up Before You Regret: Backup Awareness in Coding Agents
SummaryThis paper examines whether coding agents, which autonomously modify files, environments, and configurations, recognize when to preserve state before changing it. The authors introduce RegretBench, a benchmark of 205 executable tasks drawn from user regret moments on platforms such as X.com, GitHub, and Stack Overflow. Across six frontier models, agents completed 86.1% of tasks on average but backed up beforehand in only 15.2% of cases. The authors also propose BackUpper, a plug-in backup layer that checks each tool call, and report its results on the benchmark.
In briefData Differentiator Framework: Measuring How Much New Value a Dataset Adds
SummaryThis paper presents the Data Differentiator Framework (DDF), a method for scoring whether a candidate dataset adds information beyond what an organization already holds, and whether that gain justifies its cost. It combines a mutual-information redundancy penalty, a cost-normalized and significance-tested marginal utility score, and business-set governance inputs, run through a staged pipeline called SDSAE. Validated on a public credit-card fraud benchmark with five constructed candidate assets, only the genuinely differentiated asset passed the significance gate and received an acquire recommendation.
In briefOne Policy, Many Styles: Where to Aggregate Preferences in Humanoid Robot Control
SummaryThis paper studies steerable humanoid robot control, where one policy changes its behavioral style at runtime to match user preferences. The authors use Preference Reward Responsiveness (PRR) along with coverage and utility metrics to compare four ways of aggregating preferences: reward-level, advantage-level, loss-level, and PCGrad, across two humanoid-control frameworks. No single aggregation level performs best on all environments and criteria. Early aggregation yields a wider range of behavioral responses but more unintended changes in other objectives, while later aggregation improves control precision at some cost to responsiveness.
In briefQuery Feedback and Teaching for Stable and Certifiable Language Generation
SummaryThis paper studies language generation in the limit, asking whether a learner observing an enumeration of an unknown language can eventually produce only valid, unseen strings. It introduces certifiability, a stronger notion than stability, requiring the generator to confirm from its observations that its current output is valid. Using teaching sets from learning theory, it defines a hierarchy of teaching dimensions and characterizes stable and certifiable generation in the membership-query model, relating these to identification in the limit.
Farnam Mansouri, Alireza F. Pour, Grigoris Velegkas et al.
In briefA model-free continuous-time RL method learned from fixed discrete-time batch data, with end-to-end guarantees
SummaryThis paper studies continuous-time reinforcement learning where the state follows an unknown stochastic differential equation and only a fixed batch of discrete-time transitions is available. It proposes a model-free PhiBE policy iteration method (PhiBE-PI) that learns a single state-action critic from sampled increments and reuses the same batch throughout. Under linear function approximation, the authors give a local end-to-end suboptimality guarantee and show a bias-variance trade-off: smaller sampling intervals improve the continuous-time approximation but can increase statistical error, so finer discretization does not necessarily yield better policies.
WhyProposes a model-free continuous-time policy iteration with end-to-end suboptimality guarantees and a bias-variance trade-off, largely theoretical with limited empirical validation.
In briefD-Skill: Transferring Teacher Model Knowledge into Readable Agent Skills
SummaryThis paper introduces D-Skill, which treats a large teacher agent as a black box and compares paired rollouts to distill its behavioral advantages into a natural-language skill document. The skill is built offline and given to a smaller student at inference time, without any weight updates. Across SpreadsheetBench-Verified, ALFWorld, and chess puzzles, it closes 75%-95% of the teacher-student gap and beats one-epoch SFT baselines. The student also has 5.7x-8.3x lower inference cost per task than the teacher within the same model family.
In briefSimply Logical at Thirty: A Prolog Textbook's Path from Print to Interactive Web
SummaryThis paper traces the history of the Prolog textbook Simply Logical, from its 1994 print edition with runnable programs on a 3.5-inch diskette through several intermediate online versions to today's fully interactive online edition. It examines, from both writer and reader perspectives, how technology can affect educational materials. The authors cover the book's teaching philosophy, its web-based interactive realisation, and an authoring toolkit that supports the online edition. The paper reflects on future opportunities for interactive Prolog teaching materials.
In briefBacktracking Direct, Indirect and Total Effects: causal effect measures built by tracing back alternative scenarios instead of intervening
SummaryThe paper proposes a new way to define direct and indirect causal effects, based on backtracking counterfactuals rather than the standard cross-world interventional counterfactuals. Instead of intervening on the treatment or mediator, it uses a backtracking conditional distribution to construct an alternative exogenous state that satisfies a desired treatment-mediator constraint. This gives backtracking direct, indirect and total effects, and the population intervention effects are adapted accordingly.
In briefDiffResearch: Speeding Up Deep Research with Diffusion Language Models
SummaryThis paper presents DiffResearch, an open multi-agent deep research framework that uses a diffusion language model as the core of each stage. Keeping the scaffold, retrieval code, prompts, and evaluation fixed, swapping in a diffusion backbone made the end-to-end process roughly 6 times faster across 100 Deep Research Bench tasks, mainly by reducing planning and writing time. The authors also introduce dFast, which extracts subqueries from intermediate denoising states before generation finishes, reporting a 1.99 times median speedup for decomposition alone with no loss of report quality.
Alan-Barsag Gazzaev, Alexey Gavrilov, Mikhail Mozikov et al.
In briefHidden Messages Between Diffusion Language Model Agents: Causal Evidence for Latent Communication
SummaryThe paper studies whether two independent diffusion language model agents can pass compact hidden states to each other before decoding to text, and whether the receiver actually uses the message content. Removing or swapping the message changes accuracy substantially, for example on QASC accuracy falls from 80.8% to 46.8% with a wrong-example message. A single late transfer reaches 87.8%, outperforming four recurrent exchanges, suggesting timing of transfer may matter more than frequency.
WhyProposes hidden-state communication between diffusion LM agents, with causal ablations and substitutions showing content effects and temporal findings.
Alan-Barsag Gazzaev, Alexey Gavrilov, Sergey Muravyov et al.
In briefTropical cyclone properties encoded in sparse features of an AI weather model
SummaryThis paper uses tropical cyclones in the GraphCast AI weather model as a test case, analyzing 2,115 six-hourly timesteps from 205 cyclones. It relates sparse autoencoder activations to cyclone position, intensity, and pressure structure. Each attribute maps to multiple features, with position and pressure-structure relations partly overlapping, and this organization recurs across independent autoencoders and layers.
In briefData-Driven Glyph Positioning for Handwritten Music Synthesis
SummaryThis paper studies how to place musical symbols when generating synthetic handwritten scores for Optical Music Recognition, a field short of annotated data. The authors measure symbol spacing in about 285,000 annotated relationships and find that some hard-coded positioning rules differ from real handwriting. Replacing them with empirical means and a graph neural network reduces positioning error, and a detector pretrained on the improved synthetic data gains about 30% mAP50 after fine-tuning on 200 real staves.
In briefMinimax-optimal predictive flow matching under distribution drift
SummaryThis paper studies how to deploy a generative velocity field before current data arrive when the data distribution drifts over time. The authors propose DyProd-FM, which refits flow matching on dyadic historical prefixes and selects among fits with a sleeping Prod master, achieving dynamic regret bounds without knowing any drift budget. For fixed neural classes, matching lower bounds show both rates are optimal in T and the budgets up to logarithmic factors, without requiring convexity or curvature assumptions.
WhyGives minimax-optimal dynamic regret for nonconvex neural classes under drift without curvature assumptions, with matching lower bounds; strong theory, moderate scope.
In briefIs the refusal direction universal across low-resource languages?
SummaryThis paper tests whether the refusal direction in safety-aligned language models transfers across languages, extending PolyRefuse to Belarusian, Bashkir, and Tajik. The authors replicate the pipeline of Wang et al. (2025) on Qwen2.5-14B-Instruct. They find universality is graded: directions from safety-aligned and partially aligned languages transfer broadly, while those from strongly misaligned languages fail to reliably modulate refusal behavior.
In briefCSD: A method for transferring knowledge between models with different tokenizers
SummaryThis paper introduces Common-Support Distillation (CSD), a logits-based framework that lets a student model learn from a teacher model even when they use different tokenizers. At each aligned input chunk, CSD builds a shared support between teacher and student and matches the induced distributions with a single objective. The authors report that across four teacher-student settings, CSD achieves the highest average performance among the evaluated cross-tokenizer methods, and also outperforms continued pretraining and same-tokenizer distillation.
Mingyu Yang, Sharath Turuvekere Sreenivas, Adithyakrishna Venkatesh Hanasoge et al.
In briefReducing catastrophic forgetting when adapting language models to new domains by reusing existing skills
SummaryThis paper addresses catastrophic forgetting, where language models lose earlier abilities when post-trained on new domains. It proposes Skill Schema Transport (SST), which recasts new examples as instances of reusable procedures already in the model's skill space rather than constraining updates to stay near the old model. Across two 3B instruction-tuned backbones, SST improves average final performance on streamed domains by 2.53 to 5.75 points over sequential post-training and reduces max-drop forgetting in all evaluated settings.
In briefFig: A System That Automatically Elicits Multiagent Behaviors in Games
SummaryThe paper introduces Fig, an agentic system that studies multiagent behavior by intervening in game environments rather than only observing emergent behavior. It turns a natural-language behavior description into sub-behaviors, generates or selects social games, and applies interventions to engines, opponents, instructions, or context. Across six behaviors, models attempted the target behavior in 45.2% of episodes on average for valid scenarios.
In briefPolicy-Mask: Budget-Aware Selection of Diffusion Features for Discriminative Tasks
SummaryThis paper addresses how to pick useful intermediate features from a frozen diffusion model for tasks like segmentation, classification, and semantic correspondence. It proposes Policy-Mask, a budget-conditioned policy trained on the downstream objective that jointly selects the timestep, block, and channel subset under a fixed channel budget. The authors normalize the channel-score term by subset size to stabilize joint training. The paper reports gains over fixed extraction and feature-filter baselines at matched budgets and extends the method to diffusion transformers such as PixArt-Σ and SD3.5 MMDiT.
In briefBuilding AI research teams to autonomously discover linear attention architectures
SummaryThe paper proposes a team of persistent AI researchers, each with a different scientific discipline, who keep private research histories, share experimental evidence, and receive research opportunities through adaptive allocation. The team was used to search for linear attention architectures and outperformed GDN2 at 1.3B scale on language modeling, common-sense reasoning, and retrieval tasks. Controlled comparisons show that removing any single organizational component markedly weakens the discoveries.
WhyTeam of AI researchers with disciplinary backgrounds and adaptive allocation beats GDN2 on linear attention search, with ablations supporting each component.
In briefSANE: Automated, scorer-specific metamorphic testing for checking whether LLM judges score reliably
SummaryThis paper introduces SANE, an agent-based framework that automatically tests the reliability of LLM judges. It first identifies scorer-specific capabilities, then derives perturbation recipes and expected score relations for each, and validates the generated tests with a scorer-independent check. Across five domains, SANE produces 3.7 times as many validated test families as the Judge Reliability Harness (JRH) and detects scorer deficiencies in every domain. On faithfulness, its test-suite scores rank 11 judge models consistently with human-labeled FaithBench performance (Spearman rho = 0.852), versus 0.100 for JRH.
In briefRL-IF: Tracing How Past Training Data Affects Later Online RL
SummaryThis paper proposes RL-IF, a data attribution method for online reinforcement learning that measures how a single training sample influences a downstream target after later online updates, not just within one round. The authors give approximation guarantees under smoothness assumptions and show the optimization effect alone often approximates the full estimator well. They build RINSE, a data filtering method, which improves training efficiency and final performance over PPO and local attribution baselines on control benchmarks and an RLHF-style toxicity task.
In briefinr2action: Encoding robot episodes as neural implicit representations for recall and action
SummaryThis paper proposes a parameter-space memory framework that stores each robot episode as a separate neural implicit representation, namely a weight delta fitted to the episode's video from a shared meta-learned initialization. The delta has a fixed size regardless of episode length and is mapped by a learned projector into token embeddings used for question answering and command regression without frame retrieval. On RoboMME, the language model exceeds chance on all question types and the decoder reaches a mean R² of 0.51. Replacing the delta with matched noise reduces performance by 11.4 points.
Leann Alhashishi, Christos Margadji, Sebastian William Pattinson
In briefReal2Sem2Real: 3D Semantic Augmentation for Robot Learning from One Human RGB Video
SummaryThis paper presents Real2Sem2Real (R2Sem2R), a pipeline that turns one monocular RGB video of a human demonstration into thousands of robot demonstrations across object instances in a category. It tracks object poses with DINOv3-distilled Gaussian splats and aligns digital cousins, generated from a few photographs, to the demonstrated objects using semantic alignment. On two real robots and four tasks, vision-language-action policies trained on this data match or exceed policies trained on teleoperation data.
Matthew Strong, Ankush K. Dhawan, Julien Poffet et al.
In briefEvery learnable distribution class is also privately learnable
SummaryThe paper shows that any class of distributions learnable in total variation distance with finitely many samples, to arbitrarily small target error, is also learnable under (ε, δ)-differential privacy to arbitrarily small target error with finitely many samples. This proves a conjecture of Ashtiani. Earlier work by Bun, Kamath, Mouzakis, and Singhal had weakly refuted the conjecture under a fixed-target-error definition of learnability. The authors show the conjecture holds under the standard definition of learnability.
WhyResolves a conjecture on distribution learnability versus differential privacy, overturning a prior weak refutation with a rigorous proof of significant theoretical importance.
In briefTo Manipulate or Not: How Opinions Shift When Two LLMs Debate
SummaryThe paper sets up a controlled debate between two language models on climate-change attribution, where one role has an assigned stance and, in some conditions, is told to persuade. It compares four ordered model combinations, including GPT-6 Luna and Gemini 2.5 Flash Lite, tracking each model's explicitly reported position and its turn-by-turn behavior. Same-model pairs narrowed the opinion gap on average, while cross-model pairs widened it. The authors find that explicit reports and in-dialogue behavior capture different aspects of opinion change, and that instructing a model to persuade does not reliably increase its influence.
Mohammad Siahkamari, Phuoc Dat Nguyen, Stephen MacNeil
In briefHow much gradient truncated backpropagation loses in looped Transformers
SummaryLooped transformers reuse one block of layers many times and are often trained by backpropagating through only the last k loops. The authors measure the weight gradient lost by this truncation for every k on nine looped language models. On trained models, the lost gradient stays nearly constant as loop count grows well above k, and it falls during fine-tuning. The paper could not show that this measure predicts the training cost of truncation better than simple rules.
In briefA general topological index for symmetries and critical orbits in neural network loss landscapes
SummaryThe paper introduces a topological symmetry index based on equivariant degree theory to characterize symmetries in neural network loss landscapes and the critical orbits they produce. The framework applies to arbitrary network architectures and includes methods for degenerate critical orbits. The index detects the isotropy of critical points and quantifies their net topological contributions. Examples include softplus MLPs, Transformers, and equivariant neural networks.
In briefAn AI Agent for Managing User-Centric Open RAN Networks
SummaryThis paper addresses the challenge of forming serving clusters in user-centric Open RAN networks. The authors propose an AI agent built on a language model to solve this clustering problem. Simulation results show that the approach more than doubles the median user throughput compared with a standard network-centric architecture.
In briefDualHeads: A simple method for reward-guided few-step image generation
SummaryThis paper addresses aligning distilled few-step diffusion image generators with task goals and human preferences using only external reward evaluations and reference samples. The proposed DualHeads method adds two lightweight binary classifier heads to a frozen diffusion backbone: a reward head that learns which same-prompt samples score higher, and a reference head that separates reference from generated images. The generator maximizes a weighted sum of their logits. With rule-based and model-based rewards, the few-step generator reaches task rewards close to 40-step reinforcement learning methods while preserving image quality at no extra inference cost.
Yuanzhi Zhu, Eleftherios Tsonis, Simo Alami Chehboune et al.
In briefDepth-Guided Geometric Adaptation for Classifying Images from Very Few Examples
SummaryFew-shot image classification struggles when background clutter and similar colors make it hard to match key features with little data. The paper proposes DAFIC, which uses monocular depth to guide representation learning, episodic adaptation, and dense support-query matching. A depth-guided module refines image patches while suppressing background interference, and bidirectional patch matching reduces errors from spurious correspondences. Experiments on four benchmarks report strong performance against existing methods, supported by ablation and robustness analyses.
Zhiqi Miao, Estefan\'ia Talavera, Weichuan Zhang et al.
In briefSpeedrun leaderboard for relational foundation model pretraining (modded-nanoRT)
SummaryThe paper introduces modded-nanoRT, the first speedrun leaderboard for Relational Transformers, which ranks entries by wall-clock time to a quality target on an open, fixed problem that a single GPU can run. The authors choose the held-out database through calibration and open the leaderboard with a recipe combining typed trilinear attention, Muon, and SIGReg. The recipe reaches the target in 305.63 seconds on one NVIDIA H200, 8.11 times faster than the reference implementation, and wins 13 of 15 pairs outside the development database. The benchmark, frozen artifacts, and record runs are slated for release.
Gabriel Wedson Mendon\c{c}a de Souza Veloso, Luan Borges Teodoro Reis Sena, Denis Oliveira Correa et al.
In briefSelf-supervised tuning of simplified optimization models using detailed-model feedback
SummarySimplified (low-fidelity) optimization models are efficient to solve but systematically differ from detailed (high-fidelity) models that better capture physical constraints. The paper proposes a self-supervised framework that tunes the simplified model directly from feedback of the detailed model, using its dual variables for gradients and requiring no high-fidelity solution labels. Tested on optimal power flow, optimal gas flow, and pooling problems, the method improves objective quality and feasibility over untuned models and generally outperforms semi-supervised training.
Gonzalo E. Constante-Flores, Andr\'e H. Quisaguano-Paredes, Asha Ramanujam et al.
In briefWhen Gaussian Priors Fail: A Latent Geometry Study of RL Fine-Tuning
SummaryThis paper examines why Gaussian priors on VAE latent spaces become a limitation during reinforcement-learning fine-tuning of latent diffusion models. On the alanine-dipeptide Boltzmann distribution, a Gaussian-prior VAE collapses at beta_c = 0.224, while a mode-tracked Gaussian-mixture prior collapses at 0.00316, a 71x gap. The authors attribute this gap to one geometric mediator, d_min/(2 sigma_0), and verify the causal chain through controlled interventions.
In briefBeyond Correct Answers: How Structured Cues Affect Visual Judgments
SummaryThis paper studies how vision-language models bind attributes and relations to the correct objects in visual question answering, using structured prompts with labeled fields. Across ten models and six tasks, the authors isolate the effects of individual field labels and of auxiliary hints that preserve or swap role assignments. They find that overall accuracy can hide these differences, and that clearer hint wording improves text matching without consistently improving visual judgments. Directly selecting the relation shown in the image partly reduces errors for some hints in Qwen3 and GPT-4.1-mini.
In briefKIRA: A Children's Reading Screener with Two-Pathway Speech Analysis
SummaryThis paper presents KIRA, a full-stack reading screening system for children that combines word-level fluency scoring with phoneme-level error detection. The authors adapt pretrained word and phoneme recognizers to children's speech, reducing word error rate from 13.88% to 8.84% and phoneme error rate from 41.66% to 23.00%. The system implements Word Reading and Oral Reading Fluency tasks for K-2 screening and includes a web-based interface for teachers.
In briefUsing parallel text to improve multilingual pretraining
SummaryThis paper studies how to use the parallel structure of multilingual data in pretraining by placing each document together with its translation in the same training sequence. Compared with standard pretraining at matched token budgets, this parallel approach alone does not consistently improve general language understanding, though it builds a stronger cross-lingual foundation. The authors propose a two-stage method that first uses parallel pretraining and then standard pretraining, which consistently outperforms standard pretraining alone in both high-resource and low-resource settings.
In briefDuoPIC: Uncertainty-aware voxel-level digital twin of brain blood flow for stroke
SummaryThis paper presents DuoPIC, a digital twin of cerebral perfusion that needs no reference maps. It uses a Windkessel flow model to drive a tracer transport PDE, with a neural closure splitting each region into fast and slow pathways and a posterior quantifying uncertainty. Its voxel maps achieve AUCs of 0.745 on ISLES'17 MR and 0.743 on ISLES'24 CT for lesion detection. On held-out patients, the neural PDE cut reconstruction error by 19% and raised 95% MTT interval coverage from 88% to 94%.
Anh Duong Tran, Quang-Anh N. D., Toan Sy Pham et al.
In briefSketch2Sim: an end-to-end modular framework for hand-drawn circuit analysis
SummaryThis paper presents a system that converts photographs of hand-drawn circuit diagrams into SPICE-compatible netlists and automatically computes nodal voltages and currents. The pipeline has three stages: component and text detection, value recognition using a CRNN with CTC and OCR, and node identification via thresholding, skeletonization, and DBSCAN. The authors curated a dataset of 1,000 hand-drawn RLC circuits and report robust performance across handwriting styles, lighting, and clutter.
Mohammad Kawsar, Al Amin Islam Ridoy, Md Rakib Hossain et al.
In briefConditional Imaging: Protecting Camera Footage at Capture Yet Recoverable for Authorized Use
SummaryThe paper argues through information-theoretic analysis that static optical systems cannot simultaneously guarantee privacy and later recoverability, since information kept for reconstruction is also available to observers. It proposes conditional imaging, which changes the optical state for every captured frame, so the measurement alone conceals the scene while knowledge of the system state enables reconstruction. Experiments on a captured human-face dataset show that authorized reconstructions preserve identity, while observers without the system state cannot access visual information.
In briefWhere a stuck coding agent should spend its next inference resources, and how to measure it
SummaryThe paper introduces Banzhaf Resource Attribution (BRA), a factorial framework that measures how computation, information, and feedback each contribute to repairing failed code, holding the failed state and revision protocol fixed. It evaluates six models on 194 function-level failures from MBPP+ and HumanEval+ and 126 repository-level failures from SWE-bench Verified. Task-relevant specifications consistently improve repair, public-test feedback gives small gains, and retrieved code gives little to none. Three of four frontier models choose feedback in 95-100% of cases yet obtain lower gains than uniform random resource selection.
WhyProposes a Banzhaf attribution framework using controlled factorial experiments, revealing that agents prefer feedback yet gain less than random resource selection.
In briefHorizonComposer: Spatiotemporally Consistent Driving Video Editing with Enriched Traffic Semantics
SummaryThis paper presents HorizonComposer, an instruction-guided video editing model for simulating varied driving conditions to stress-test autonomous driving systems. It combines real and synthetic video data, uses dense pixel-level rewards from offline reinforcement learning to steer generation, and introduces thinking frames to smooth transitions between image and video conditioning. The authors report better preservation of multi-actor dynamics, higher user preference scores, and gains in downstream road line segmentation and multi-object detection.
In briefWorldStream: Streaming Interactive 3D Worlds with Hierarchical Planning and Persistent State
SummaryWorldStream is an open-ended 3D world model that extends playable terrain beyond a predefined map, acts on free-form language within the running world, and restores visited places with their edits. A hierarchical planner produces macro and micro structured decisions, which deterministic local code compiles into terrain, roads, settlements, and vegetation. A single persistent world record stores plans, seeds, object states, and edits, so returning tiles are recompiled from it.
In briefDecision language models: a cheaper, faster way to judge clinical AI outputs
SummaryThis paper compares Jev, a decision language model that returns structured judgments, with frontier and open-weight LLM judges for evaluating clinical AI outputs. Across four use cases, Jev's agreement with physicians matched or exceeded that of many LLM judges, including some frontier models. Frontier judges were reported to cost 73 to 294 times more and take 7 to 57 times longer than Jev. The authors suggest Jev may enable efficient clinical AI evaluation at scale.
In briefA parallel treebank mapping sentence structure across five Arabic city dialects
SummaryThis paper introduces MADARTB, the first large-scale parallel dependency treebank covering five Arabic city dialects (Beirut, Cairo, Doha, Rabat, Tunis) and extending an existing MSA version. It contains 5,000 sentences annotated with the CATiB scheme. The authors find that global syntactic profiles are highly similar across dialects while local structures often differ. A supervised parser reaches 87.5% LAS, outperforming the best of four LLMs at 85.7%. The annotations and models are publicly released.
Noor Abo Mokh, Mohamed Adel, Jamila El-Gizuli et al.
In briefRegister Anything Model: One Model for Aligning 3D Point Clouds Across Many Domains
SummaryThis paper presents Register Anything Model (RAM), a single model that aligns 3D point clouds from indoor, outdoor, and object domains by estimating 6-DoF transformations. It introduces a geometry-aware normalization strategy that standardizes point clouds despite differences in scale, noise, and density, followed by a hierarchical coarse-to-fine design with an efficient geometric transformer. Across 12 benchmarks, the authors report state-of-the-art inlier ratio and registration recall, outperforming models trained separately on each domain.
In briefLIFT: Scoped Constraints for Testing and Training Instruction Following
SummaryThis paper introduces Lift-Bench, an instruction-following benchmark in which models draft a document from a topic and a structured outline, with constraints limited to specific sections. It includes both programmatically verifiable and judged constraints, plus a training counterpart, Lift-Train. The authors report that frontier models struggle on the harder subsets, and that training on Lift-Train is less prone to reward hacking than training on IFTrain.
Eran Goldstein, Yuval Belfer, Yonatan Belinkov et al.
In briefBrittle by Design? Robust Mission Coordination via Multi-Scenario Reinforcement Learning
SummaryThis paper proposes a method for coordinating missions in Systems of Systems by combining high-fidelity digital mission models with Reinforcement Learning for adaptive task allocation. Tested on an aerial firefighting mission across three scenarios, trained coordinators significantly outperformed random and heuristic baselines. Coordinators trained on a single scenario could degrade when transferred to harder conditions, while a coordinator trained on both scenarios concurrently performed best within the same training budget.
In briefAdding More Tasks to Better Agent Harnesses: Successive Scale Racing for Evolution
SummaryThis paper studies evolving an AI agent's harness, its instructions, tools, and control flow, using its past task experiences. It finds that existing methods perform worse as the task pool grows, because evidence is selected poorly during optimization. The authors propose Successive Scale Racing (SSR), which searches for the best evidence combinations in each round and races candidate harnesses under a small compute budget. The paper reports state-of-the-art results across six configurations on three benchmarks, a pass rate rising from 64.3% to 68.3% as tasks grow from 40 to 200, and 60% fewer validation executions.
In briefA Mechanistic Study of Automated Failure Attribution in LLM Agent Systems
SummaryThis paper examines why language models struggle to automatically identify which agent step caused an incorrect outcome in multi-agent systems. Using linear probes on internal activations, the authors find a "knows-but-won't-say" gap, where models encode enough information to attribute failures correctly but do not explicitly output the responsible step. They propose grafting, a manual intervention that uses this latent knowledge without requiring ground-truth labels at inference, and report that it consistently outperforms direct prompting across several LLMs and benchmarks.
In briefSAIL: A learnable generalisation of spatial autocorrelation for graph data
SummaryThis paper introduces SAIL (Spatial Autocorrelation Index Learner), a differentiable framework that extends classical spatial autocorrelation statistics such as Moran's I and Geary's C into a learnable operator. It jointly learns feature projections and inter-node neighbourhood affinities to produce interpretable local and global spatial coherence descriptors. Experiments on synthetic data and large biomedical graphs are reported, with applications in vision, geospatial analysis and biomedicine.
In briefToward More Realistic Models of Change: Causality for Decisions and Dynamical Systems
SummaryThe paper notes that statistical learning often fails to generalize under distribution shift, and argues that causal models, which capture mechanisms of change, offer a remedy. The first part studies algorithmic recourse, proposing a more realistic cost function, a causal reformulation, and a temporal dimension for policy evaluation. The second part extends causal interventions and effects to discrete and continuous stochastic processes, linking their equilibria to Structural Causal Models as compressed representations of temporal dynamics.
In briefJUICE: One model that generates both images and text in pretrained encoder spaces
SummaryThe paper introduces JUICE, a single flow matching Transformer that models images and text jointly over frozen pretrained encoder representations, giving each modality its own noise level. Which modalities are clean or generated determines the task, covering image generation, captioning, visual question answering, and joint generation. On mazes with multiple valid solutions, coupled sampling raises joint success from 54.6% to 93.7%.
In briefDepth Separation for Neural Networks Under Gaussian Measure
SummaryThis paper revisits depth separation for neural networks, studying L² approximation of radial functions under the standard Gaussian measure in d dimensions. It shows a radial function that a depth-3 ReLU network of polynomial size in d can approximate to small error, while depth-2 networks need size at least exponential in d unless their weights are subexponentially bounded. A general result is proved: radial functions with substantial mass in high Hermite frequencies cannot be well approximated by small-width, bounded-weight depth-2 networks. All theorems are verified in Lean 4.
Cynthia Bortolotto, Francisco Ganacim, Roberto I. Oliveira
In briefWhere do legal factors live inside large language models?
SummaryThis paper asks whether LLMs form internal representations for all 26 CATO legal factors. The authors test three open models, Qwen3-1.7B, Qwen3-8B and Gemma-2 2B, using linear decoding and sparse autoencoders. All 26 factors are linearly decodable in all three models, and single-feature sparse readouts reach best-layer test AUCs from 0.86 to 1.00.
